High Performance Computing and Modern Science Prof. Dr. Thomas Ludwig

Transcription

1 High Performance Computing and Modern Science Prof. Dr. Thomas Ludwig German Climate Computing Centre Hamburg Universität Hamburg Department of Informatics Scientific Computing

2 Abstract High Performance Computing (HPC) became an instrument for many modern sciences. In particular numerical simulations are based on parallel computing in HPC systems. Beside experiment and theory they are considered as the third pillar of knowledge gaining. However, HPC became very expensive over the last ten year. In particular, modern systems ask for huge amounts of electrical power such that the possible research starts to be limited by the research institution s budget for power. Moreover, programming these systems that exhibit massiv parallelism in their components requires scientists knowledgeable in software engineering. Finally, there is now a new challenge for our researchers: the aggregation of measured and generated digital data leads to new epistemological concepts. It is usually called data intensive science or the Fourth Paradigm. Available data becomes the primary product for new and unique insights. The talk will discuss current challenges in HPC and data intensive science and will highlight the potential development in the field until the end of the decade. Beyond 2020 the development in computer technology is difficult to predict - modern science and modern life ask however for a sustained progress. 2

3 Computational Performance x this is ONE floating point operation (FLOP) if you need 100 seconds to calculate this your computational performance is 1/100 Flops 3

4 Computers and Performance The first computer in 1941 (Zuse s Z3) made 0.3 Flops The No. 1 computer in Nov 2013 (Tianhe-2) makes 33,862,700,000,000,000 Flops 33x10 15 Flops 33 quadrillion Flops 33 PetaFlops 4

5 Computers and Science & Engineering Zuse s Z3 Fluid dynamics and wing profile optimizations ENIAC (1946, 385 Flops!) Artillery firing tables Simulate hydrogen bomb (John von Neuman) First numerical weather forecast (John von Neuman, 1950) Tianhe-2 Fluid dynamics Cosmology Business Opinion Analysis Security e-government Cloud 5

6 TOP500 Leistung factor of 1,000 in 12 years notebook years 6

7 1PFlops 1TFlops 1GFlops 1MFlops KFlops 1Flops 0,001Flops 7

8 1PFlops 1TFlops 1GFlops 1MFlops KFlops 1Flops Zuse s Z3 0.3 Flops (1941) 0,001Flops 8

9 1PFlops 1TFlops 1GFlops 1MFlops KFlops 1Flops Zuse s Z3 0.3 Flops (1941) 0,001Flops 9

10 1PFlops 1TFlops 1GFlops 1MFlops KFlops 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 10

11 1PFlops 1TFlops 1GFlops 1MFlops KFlops Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 11

12 1PFlops 1TFlops 1GFlops 1MFlops KFlops Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 12

13 1PFlops 1TFlops 1GFlops 1MFlops KFlops EC KFlops (1972) Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 13

14 1ZFlops 1EFlops 1PFlops 1TFlops 1GFlops MFlops 1KFlops EC KFlops (1972) Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 14

15 1ZFlops 1EFlops 1PFlops 1TFlops 1GFlops MFlops 1KFlops EC KFlops (1972) Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 15

16 1ZFlops 1EFlops 1PFlops 1TFlops 1GFlops (?) 1MFlops 1KFlops EC KFlops (1972) Glushkov s KIEV 3 KFlops (1958) 1Flops Bruk s M-1 20 Flops (1950) Zuse s Z3 0.3 Flops (1941) 0,001Flops 16

17 Increase in Computational Performance from 0.3 Flops in 1941 to 33x10 15 Flops in 2013 as constant exponential growth means factor every 12.5 years for already more than 70 years 17

18 Exponential Acceleration 18

19 A 1941 Ford 19

20 Geneva Bourg-en-Bresse 100 km in 1 hour 20

21 21

22 A Ford as Fast as Light 1,079,252,849 km/h will be reached in

23 The Universe and the 2013 Ford (Petascale) by the way: Tianhe means Milky Way (5min for our Ford) 9 years 90 billion lightyears 23

24 The Universe and the 2020 Ford (Exascale) 2 months 90 billion lightyears 24

25 The Universe and the 2032 Ford (Zetascale) 2.5 h 90 billion lightyears 25

26 Energy Consumption A killer argument against high performance computing today and tomorrow Tianhe-2 uses 17 MW May cost 17 million Euro per year for power Same problem with most compute HPC centers But 26

27 Time will tell... Currently energy efficiency increases by a factor of 10 every 7 years years of DKRZ

28 What Type of Science? Basically we do numerical simulations Climate and weather simulation Cloud simulation Cosmology The origin of the universe Particle physics Proteine folding Individualized medical treatment Brain simulation 28

29 Orders of Magnitude Tianhe-2 in x10 15 Flops Flops/neuron Human brain 100x10 9 neurons Exaflops in x10 18 Flops 10 million Flops/neuron Zetaflops in x10 21 Flops 10 billion Flops/neuron 29

30 d a t a d e l u g e

31 The Fourth Paradigm Refer to Jim Gray et al.: The Fourth Paradigm Data-Intensive Scientific Discovery Data are not consumed by the ideas and innovations they spark but are an endless fuel for creativity Harnessing the Power of Digital Data for Science and Society by the Interagency Working Group on Digital Data January

32 Climate Science and Data Climate research was and is dependent on data Keep data for long term in order to validate models in the future Keep all model variables because we do not know what will be necessary in future analyses Use a fine global grid in order to select later local regions... DKRZ archived around 8 PB on tape in 2013 This is all output data of simulation runs 32

33 Data Generation DKRZ (115 TFlop/s, 26 TByte main memory) produces an estimated data transfer mem<->disk 5-10 GB/s ( TB/day) ca. 100 TB/day are saved for further inspection ca. 20 TB/day are archived to tape 33

34 Example CMIP5 The 5th Coupled Model Intercomparison Project Provides key input for the next IPCC report (5 th AR) Intergovernmental Panel on Climate Change ~20 modeling centers around the world (DKRZ being one of the biggest) Produces 10s of PBytes of output data from ~60 experiments ( digital born data ) Data are produced without knowing all applications beforehand and these data are stored and archived for interdisciplinary utilization by yet unknown researchers 34

35 Data creation Example Workflow CMIP5 German contribution produced at DKRZ Used ½ year of its compute capacity Data archival 600 TB raw + 60 TB with quality control Data dissemination Will be heavily used during the next 5 years Expected CMIP6 data volume: factor higher 35

36 The Next 10 Years [TB] ,1 0,01 0,001 AR AR AR AR AR AR

37 Future Computer and Storage Next generation climate computer 2-3 PFlop/s 45 PByte on disk An estimated ½ EByte as tape library capacity Storage: 30-50% of investment and energy costs Currently it is more like 10% Storage 50% of the overall complexity? 37

38 Where will this leads us to? 38

39 Where will this lead us to? Unprecedented science E.g. life sciences: genetic design, articificial DNA Unprecedented products E.g. medical science: prosthesis, medical treatments Unprecedented problems Ethical question will be undiscussed Societal considerations will be undiscussed Legal aspects will be undiscussed 39

40 Technological Singularity The technological singularity, or simply the singularity, is a hypothetical moment in time when artificial intelligence will have progressed to the point of a greater-than-human intelligence, radically changing civilization, and perhaps human nature. Because the capabilities of such an intelligence may be difficult for a human to comprehend, the technological singularity is often seen as an occurrence beyond which the future course of human history is unpredictable or even unfathomable. (Wikipedia) Singularity mentioned first in 1958 by John von Neumann See also his book The Computer and the Brain See Ray Kurzweil s non-fiction book: The Singularity is Near 40